Graph data preprocessing method, system and equipment and medium
By removing perturbing edges in graph data and using the depth map mutual information model DGI to learn mutual information of graphs, the problem that the existing technology cannot effectively remove the tiny perturbing structure in graph data is solved, and the improvement of the adversarial attack training effect and model reliance is achieved. Augmented strength.
Patent Information
- Application Number
- CN202411600754.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-11
- Publication Date
- 2025-02-18
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing technology cannot effectively remove the tiny perturbation structures in the graph data, resulting in the graph neural network being fragile in the face of malicious adversarial attacks, with poor preprocessing effects, affecting the training effect of adversarial attacks.
By removing multiple perturbing edges of the pending graph data, a rough map is generated; then randomly select some edges from multiple perturbing edges to add them to the coarse graph to be predicted; use the depth map mutual information model DGI to learn the mutual information of the graph , and determine whether the contamination is less than the threshold value by the graph. If not, continue to add disturbed edges and repeat the judgment until a clean graph is obtained.
By gradually purifying the graph data, the tiny perturbation structure in the graph data can be effectively removed, the adversarial attack training effect of the graph neural network can be improved, and the robustness of the model can be enhanced.
Smart Images
Figure CN119474870A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of graph representation learning, and specifically relates to a graph data preprocessing method, system, device and medium. Background Art
[0002] Graph data is a ubiquitous data structure that can represent various entities and their complex interconnected relationships. Graph data can reflect various real-world networks, including protein networks, transportation networks, social networks, etc. At present, graph neural networks (GNNs) have become a powerful tool for effectively analyzing graphs due to their significant performance in tasks such as node classification, link prediction, and graph classification. In particular, after the introduction of unsupervised learning in graph neural networks, it is possible to learn without relying on labels while hardly changing the representation learning quality of GNNs.
[0003] However, many studies have shown that when graph neural networks are attacked by malicious adversarial attacks, they show significant vulnerability. In other words, when there is a slight perturbation (contaminated graph data) in the structure of the graph data input into the graph neural network, it is possible for the graph neural network to produce incorrect prediction results. Therefore, it is necessary to preprocess the graph data (attacked graph) input into the graph neural network, identify the perturbed structure in the contaminated graph data, and remove the perturbed structure to improve model performance.
[0004] At present, some studies have adopted similarity-based detection methods to determine whether a structure has been perturbed, that is, deleting edges between nodes with very low node representation similarity. However, these methods perform poorly in defending against adversarial attacks, because adversarial attacks are more likely to connect nodes that are usually far away from each other, and perturbed edges may cause the model to learn similar representations for nodes that are not originally adjacent, making it impossible to effectively identify small perturbation structures, resulting in poor preprocessing of contaminated graph data and affecting the adversarial attack training effect of graph neural networks. Summary of the invention
[0005] In order to overcome the deficiency of the prior art that it is unable to remove small disturbance structures in graph data, the present invention provides a graph data preprocessing method, comprising the following steps:
[0006] Remove multiple disturbed edges of the graph data to be processed to obtain a rough graph;
[0007] Some edges are randomly selected from multiple perturbation edges and added to the coarse graph to generate a graph to be predicted; the deep graph mutual information model DGI is used to learn the mutual information between the coarse graph and the spatial representation of the coarse graph, as well as the mutual information between the graph to be predicted and the spatial representation of the graph to be predicted; the graph contamination degree is used to determine whether the difference between the two mutual information is less than a threshold value, if not, multiple edges are selected from the remaining perturbation edges and added to the graph to be predicted to generate a new graph to be predicted, and it is determined again whether the difference between the mutual information of the new graph to be predicted and its spatial representation and the mutual information between the coarse graph and the spatial representation of the coarse graph is less than the set threshold, if so, a clean graph is obtained.
[0008] Preferably, removing a plurality of disturbed edges of the graph data to be processed to obtain a rough graph comprises the following steps:
[0009] Obtain the overall embedding representation of the subgraph of the graph data to be processed and the embedding representation of node i, and obtain the anomaly score of node i of the graph data based on the overall embedding representation of the subgraph and the embedding representation of node i;
[0010] The anomaly scores of the edges of the graph data are derived according to the anomaly scores of the graph data nodes i, the anomaly scores of the edges are sorted, and the first k edges are removed to obtain a rough graph.
[0011] Preferably, the step of obtaining the abnormality score of the graph data node i based on the overall embedding representation of the subgraph and the embedding representation of the node i comprises the following steps:
[0012] Randomly select a node i in the graph data to be processed, use the restarted random walk algorithm RWR to perform a random walk on the node i, combine the nodes and edges of the walk as positive samples, and represent the collected positive samples as a subgraph;
[0013] Use the GCN model to learn the subgraph embedding representation of node i;
[0014] Use a deep neural network DNN to map node i to the same embedding space as the overall embedding representation of the subgraph to obtain the embedding representation of node i;
[0015] According to the overall embedding representation of the subgraph and the embedding representation of node i, the abnormality score of the graph data node is obtained in combination with the bilinear scoring function.
[0016] Preferably, the anomaly score of the edge of the graph data is derived according to the anomaly score of the graph data node, and is obtained by the following formula:
[0017]
[0018] In the formula, S i is the abnormal score of node i, S j is the abnormal score of node j, A i，j=1 means there is an edge between node i and node j, A i，j =0 means there is no edge between node i and node j.
[0019] Preferably, the subgraph embedding representation of node i is learned by using the GCN model, and is performed by the following formula:
[0020]
[0021] In the formula, k represents the kth node of the subgraph, K represents the number of subgraph nodes, and H k represents the representation of the k-th node.
[0022] Preferably, the embedded representation of the node i is obtained by the following formula:
[0023]
[0024] In the formula, h i (l-1) represents the representation of node i at layer l-1, W (l-1) represents the parameters of the l-1 layer, is the ReLU activation function, h i Table 1 shows the overall embedding representation of the subgraph of node i.
[0025] Preferably, the graph contamination degree is:
[0026] GPD(θ)=I(G,f θ (g))-I(G′,f θ (G′));
[0027] In the formula, G is the rough image, G' is the image to be predicted, I(G,f θ (G)) is G and f θ (G) The mutual information between θ (G) is the spatial representation of the rough graph, I(G', f θ (G')) is G' and f θ (G'), f θ (G') is the spatial representation of the image to be predicted, and GPD(θ) represents the difference between the mutual information of the coarse image and the image to be predicted.
[0028] The present invention also provides a graph data preprocessing system, comprising:
[0029] A rough graph acquisition module is used to remove multiple disturbance edges of the graph data to be processed to obtain a rough graph;
[0030] The graph purification module randomly selects some edges from multiple perturbation edges and adds them to the coarse graph to generate a graph to be predicted; uses the deep graph mutual information model DGI to learn the mutual information between the coarse graph and the spatial representation of the coarse graph, and learns the mutual information between the graph to be predicted and the spatial representation of the graph to be predicted; uses the graph contamination degree to determine whether the difference between the two mutual information is less than a threshold, if not, selects multiple edges from the remaining perturbation edges and adds them to the graph to be predicted to generate a new graph to be predicted, and again determines whether the difference between the mutual information of the new graph to be predicted and its spatial representation and the mutual information between the coarse graph and the spatial representation of the coarse graph is less than a set threshold, if so, a clean graph is obtained.
[0031] The present invention also provides a computer device, including a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the graph data preprocessing method.
[0032] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the graph data preprocessing method.
[0033] The graph data preprocessing method provided by the present invention has the following beneficial effects:
[0034] The present invention can obtain a coarse graph by removing most of the disturbed edges in the graph data; can generate a graph to be predicted by adding perturbation edges to the coarse graph multiple times; can obtain the mutual information between the spatial representation of the coarse graph and the spatial representation of the coarse graph, and the mutual information between the graph to be predicted and the spatial representation of the graph to be predicted by using DGI; can purify the graph data by adding a perturbation edge each time and using the graph contamination degree to judge the difference between the two mutual information once; through the process of continuous iterative purification, can gradually restore the accurate node representation, remove the tiny perturbation structure in the graph data, and finally obtain clean graph data; the obtained clean graph data can improve the adversarial attack training effect of the graph neural network. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] In order to more clearly illustrate the embodiment of the present invention and its design scheme, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0036] Figure 1 This is a flowchart of a graph data preprocessing method according to an embodiment of the present invention. DETAILED DESCRIPTION
[0037] In order to enable those skilled in the art to better understand the technical solution of the present invention and implement it, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the scope of protection of the present invention.
[0038] In the description of the present invention, it is to be understood that the terms “center”, “longitudinal”, “lateral”, “length”, “width”, “thickness”, “up”, “down”, “front”, “back”, “left”, “right”, “vertical”, “horizontal”, “top”, “bottom”, “inside”, “outside”, “axial”, “radial”, “circumferential”, etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the technical solutions of the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.
[0039] In addition, the terms "first", "second", etc. are used for descriptive purposes only and are not to be understood as indicating or implying relative importance. In the description of the present invention, it should be noted that, unless otherwise clearly specified or limited, the terms "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances. In the description of the present invention, unless otherwise specified, "plurality" means two or more, which will not be described in detail here.
[0040] Example
[0041] The present invention provides a graph data preprocessing method, specifically as follows Figure 1 As shown, the following steps are included:
[0042] Step 1: Obtain the overall embedding representation of the subgraph of the graph data and the embedding representation of node i, and obtain the anomaly score of node i of the graph data based on the overall embedding representation of the subgraph and the embedding representation of node i. Specifically, the steps are as follows:
[0043] (1) Randomly select a node i in the graph data to be processed, use the restarted random walk algorithm RWR to perform a random walk on the node i, combine the nodes and edges of the walk as positive samples, and represent the collected positive samples as a subgraph.
[0044] (2) The GCN model is used to learn the subgraph embedding representation of node i, as follows:
[0045] First, the present invention needs to learn the embedded representation of the subgraph G. Since the graph convolutional neural network GCN (Graph Convolutional Network) can aggregate information from nodes in the local subgraph and transform high-dimensional attributes into low-dimensional embedding space, the present invention adopts an efficient GCN model as a model for learning node embedding. The embedding formula of the subgraph G can be expressed as:
[0046]
[0047] In the formula, is the ReLU activation function, D i represents the degree of node i, A i represents the adjacency matrix of the subgraph to which node i belongs, l represents the number of layers, represents the representation of the i-node at layer l-1, W (l-1) Represents the parameters of the l-1 layer.
[0048] Next, the present invention applies an average pooling function as a readout function to obtain the overall embedding representation of the subgraph G
[0049]
[0050] In the formula, k represents the kth node of the subgraph, K represents the number of subgraph nodes, and H k represents the representation of the k-th node.
[0051] (3) Use a deep neural network DNN to map node i to the same embedding space as the overall embedding representation of the subgraph to obtain the embedding representation of node i.
[0052] The present invention uses a deep neural network DNN (Deep Neural Networks, DNN) to calculate the representation of node i, which maps node i to The same embedding space, specifically, the embedding representation of node i is as follows:
[0053]
[0054] In the formula, h i (l-1) represents the representation of node i at layer l-1, W (l-1) represents the parameters of the l-1 layer, is the ReLU activation function, h i Table 1 shows the overall embedding representation of the subgraph of node i.
[0055] (3) Based on the overall embedding representation of the subgraph and the embedding representation of node i, the anomaly score of the graph data node is obtained in combination with a bilinear scoring function.
[0056] The present invention adopts a bilinear scoring function to obtain the abnormality score of the sample pair, and its specific formula is as follows:
[0057]
[0058] In the formula, σ represents the sigmoid activation function, Represents the subgraph G i The overall embedding representation of , W represents the parameters that can be learned, and h i represents the embedding representation of node i.
[0059] Step 2: Derived the anomaly score of the edge of the graph data according to the anomaly score of the graph data node i, sorted the anomaly scores of the edge, and removed the first k edges to obtain a rough graph.
[0060] The present invention first needs to obtain a method to exclude most of the perturbed edges, and then gradually refine them. Adversarial attacks usually connect nodes that are far away to maximize the homogeneity of the graph data, which usually results in a higher anomaly score for the two connected endpoints (nodes). Therefore, if the detected endpoint of an edge has a high anomaly score, this indicates that the probability of the edge being perturbed is high. In actual operation, the present invention relies on the anomaly score of the endpoint to derive the anomaly matrix of the edge, thereby obtaining the anomaly score of the edge, and the anomaly matrix is specifically shown as follows:
[0061]
[0062] In the formula, S i is the abnormal score of node i, S j is the abnormal score of node j, A i，j =1 means there is an edge between node i and node j, A i，j =0 means there is no edge between node i and node j.
[0063] Afterwards, the edges are sorted in descending order of anomaly scores and the top k most anomaly edges are removed to generate a coarse graph G.
[0064] Step 3: Randomly select some edges from multiple perturbed edges and add them to the coarse graph to generate a graph to be predicted; use the deep graph mutual information model DGI (Deep Graph Infomax) to learn the mutual information between the coarse graph and the spatial representation of the coarse graph, and learn the mutual information between the graph to be predicted and the spatial representation of the graph to be predicted; use the graph contamination degree to determine whether the difference between the two mutual information is less than a threshold, if not, select multiple edges from the remaining perturbed edges and add them to the graph to be predicted to generate a new graph to be predicted (gradually purified coarse graph), and again determine whether the difference between the mutual information of the new graph to be predicted and its spatial representation and the mutual information between the coarse graph and the spatial representation of the coarse graph is less than the set threshold, if so, a clean graph is obtained.
[0065] The present invention proposes to determine whether an edge is perturbed by measuring the change in mutual information. Intuitively, the goal of adversarial attacks is to destroy representations, thereby reducing the performance of downstream tasks. This shows that the representation learned from the graph to be predicted cannot fully reflect the graph containing adversarial edges. Therefore, the present invention proposes graph contamination as a measure to quantify the degree of contamination in graph data. The formula for graph contamination is as follows:
[0066] GPD(θ)=I(G,f θ (g))-I(G′,f θ (G′));
[0067] In the formula, G is the rough image, G' is the image to be predicted, I(G,f θ (G)) is G and f θ (G) The mutual information between θ (G) is the spatial representation of the rough graph, I(G', f θ (G')) is G' and f θ (G'), f θ (G') is the spatial representation of the graph to be predicted, and GPD(θ) represents the difference between the mutual information of the coarse graph and the graph to be predicted, that is, the graph contamination.
[0068] A lower GPD(θ) value indicates that the inspected graph data contains higher-risk perturbation edges. By continuously fine-tuning and purifying the rough graph based on the GPD(θ) value, the degree of contamination of the graph data G′ during the purification process can be determined. When the number of true edges in the purified graph data reaches the set threshold, the target clean graph is obtained.
[0069] In order to verify the performance of graph data processed by the method of the present invention, first, the present invention conducts experiments on three benchmark data sets, specifically selecting the network data sets Cora, Citeseer and Pubmed, which are citation network data sets, where nodes represent documents and edges represent citation links between them. In Cora and Citeseer, features are binary indicators that reflect whether specific words exist in each document. In Pubmed, features are based on TF-IDF (term frequency-inverse document frequency) values, capturing the importance of terms in documents. The edges in these data sets are created based on citation relationships, where a directed edge from one node to another indicates that the first document cites the second document. Detailed information on the features of the data sets is shown in Table 1.
[0070] Secondly, in order to evaluate the effectiveness of the method of the present invention, the present invention will use the following seven benchmark methods and the method of the present invention to purify the graph data in the above three data sets. The seven benchmark methods are as follows:
[0071] Table 1 Dataset characteristics
[0072] Dataset name Number of nodes Number of edges Number of features Number of categories Cora 2，708 5，429 1，433 7 Citeseer 3，327 4，732 3，703 6 Pubmed 19.717 44.338 500 3
[0073] Projected Gradient Descent (PGD): uses the mutual information maximization principle to perform unsupervised node representation learning.
[0074] JACCARD Similarity Method: Based on the JACCARD similarity of the representations, edges with high distances are removed.
[0075] Deep Anomaly Detection on Attributed Networks: Dominant: Uses deep autoencoders to detect perturbation edges.
[0076] Contrastive self-supervised learning framework for anomaly detection on attributed networks (CoLA): captures the relationship between each node and its neighborhood substructure in an unsupervised manner to learn anomalies and identify perturbation edges.
[0077] ANEMONE (ANomaly dEtection frameworkwith Multi-scale contrastive lEarning): A method based on contrastive learning is proposed to learn abnormal nodes.
[0078] Sub-CR (Subgraphbased multi-view self-supervised framework which contains a Contrastive-based module and a Reconstruction-based module): A dual-view contrastive learning module is proposed.
[0079] GRADATE (GRaph Anomaly Detection framework via multi-scale contrastive learning networks with newly added subgraph-subgraph contrast and Augmented view): A framework for graph anomaly detection via multi-scale contrastive learning networks with newly added subgraph-subgraph contrast and Augmented view. Figure 1 Sub-image comparison and enhanced view, through sub Figure 1 Sub-image comparison, sub- Figure 1 Node comparison and node-node comparison to detect anomalies.
[0080] Again, the above seven benchmark methods and the method of the present invention are used to purify the graph data in the three data sets, and then the purified graph data is used to perform attack training on PGD (PGD is an attack method, not a defensive one. It uses convex relaxation optimization to perform gradient-based adversarial attacks on discrete graph data. The attack principle includes calculating the gradient of the input data and adding perturbations in the gradient direction. Through multiple iterations, adversarial samples are gradually constructed, and then training representations are performed on the perturbed graph structure). The performance of the purified graph data is judged based on the training effect. The results are shown below, see Table 2 for details.
[0081] Table 2 Performance of graph data processed by different methods
[0082]
[0083] 1) Regardless of the perturbation ratio, the method of the present invention is always better than the seven benchmark methods. The reason is that the data processed by the seven benchmark methods is equivalent to the rough graph obtained by the present invention, and the method of the present invention further identifies structures containing more "true edges" based on the rough graph, thereby achieving better classification results than the seven benchmark methods.
[0084] 2) PGD performs worse than other algorithms in most cases. This indicates that the representations learned by PGD lack robustness, highlighting the necessity and importance of exploring other algorithms to learn representations with better robust performance.
[0085] 3) JACCARD and Dominant methods also perform poorly in defending against PGD attacks. This is mainly because these methods focus on structural errors rather than abnormal patterns during training, making them more sensitive to the impact of perturbed edges. In contrast, the method of the present invention avoids directly relying on structural information to determine whether an edge is abnormal, thereby performing better in defending against attacks.
[0086] 4) As the perturbation increases, the performance of our method on all datasets decreases much more slowly, indicating that the method of first selecting a coarse graph and then fine-tuning the model based on different graphs is less sensitive to perturbations.
[0087] 5) CoLA, Sub-CR, and GRADATE are three newer algorithms that use graph contrastive learning to detect graph anomalies. These methods show strong robustness on most datasets. However, they face a common problem: the obtained graph still contains some "false" edges, which may interfere with representation learning, resulting in inaccurate representation, thus affecting the performance of downstream tasks.
[0088] The present invention also provides a graph data preprocessing system, comprising:
[0089] A rough graph acquisition module is used to remove multiple disturbance edges of the graph data to be processed to obtain a rough graph;
[0090] The graph purification module randomly selects some edges from multiple perturbation edges and adds them to the coarse graph to generate a graph to be predicted; uses the deep graph mutual information model DGI to learn the mutual information between the coarse graph and the spatial representation of the coarse graph, and learns the mutual information between the graph to be predicted and the spatial representation of the graph to be predicted; uses the graph contamination degree to determine whether the difference between the two mutual information is less than a threshold, if not, selects multiple edges from the remaining perturbation edges and adds them to the graph to be predicted to generate a new graph to be predicted, and again determines whether the difference between the mutual information of the new graph to be predicted and its spatial representation and the mutual information between the coarse graph and the spatial representation of the coarse graph is less than a set threshold, if so, a clean graph is obtained.
[0091] The present invention also provides a computer device, including a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the graph data preprocessing method.
[0092] The present invention also provides a computer-readable storage medium, which stores a computer program. The computer program is suitable for being loaded by a processor to execute the graph data preprocessing method.
[0093] The embodiments described above are only preferred specific implementation modes of the present invention, and the protection scope of the present invention is not limited thereto. Any simple changes or equivalent replacements of the technical solutions that can be obviously obtained by any technician familiar with the field within the technical scope disclosed in the present invention belong to the protection scope of the present invention.
Claims
1. A graph data preprocessing method, characterized in that: The steps include: Remove multiple disturbed edges of the graph data to be processed to obtain a rough graph; Randomly select some edges from multiple perturbed edges and add them to the coarse graph to generate the graph to be predicted; use the deep graph mutual information model DGI to learn the mutual information between the coarse graph and the spatial representation of the coarse graph, and learn the mutual information between the graph to be predicted and the spatial representation of the graph to be predicted; The graph contamination degree is used to determine whether the difference between the two mutual information is less than a threshold. If not, multiple edges are selected from the remaining perturbation edges and added to the graph to be predicted to generate a new graph to be predicted, and it is determined again whether the difference between the mutual information of the new graph to be predicted and its spatial representation and the mutual information between the coarse graph and the spatial representation of the coarse graph is less than a set threshold. If so, a clean graph is obtained.
2. The graph data preprocessing method according to claim 1, characterized in that: The method of removing multiple disturbed edges of the graph data to be processed to obtain a rough graph includes the following steps: Obtain the overall embedding representation of the subgraph of the graph data to be processed and the embedding representation of node i, and obtain the anomaly score of node i of the graph data based on the overall embedding representation of the subgraph and the embedding representation of node i; The anomaly scores of the edges of the graph data are derived according to the anomaly scores of the graph data nodes i, the anomaly scores of the edges are sorted, and the first k edges are removed to obtain a rough graph.
3. The graph data preprocessing method according to claim 2, characterized in that: The method of obtaining the abnormality score of the graph data node i based on the overall embedding representation of the subgraph and the embedding representation of the node i comprises the following steps: Randomly select a node i in the graph data to be processed, use the restarted random walk algorithm RWR to perform a random walk on the node i, combine the nodes and edges of the walk as positive samples, and represent the collected positive samples as a subgraph; The GCN model is used to learn the overall embedding representation of the subgraph of node i; Use a deep neural network DNN to map node i to the same embedding space as the overall embedding representation of the subgraph to obtain the embedding representation of node i; According to the overall embedding representation of the subgraph and the embedding representation of node i, the abnormality score of the graph data node is obtained in combination with the bilinear scoring function.
4. The graph data preprocessing method according to claim 2, characterized in that: The anomaly score of the edge of the graph data is derived according to the anomaly score of the graph data node, and is obtained by the following formula: In the formula, S i is the abnormal score of node i, S j is the abnormal score of node j, A i，j =1 means there is an edge between node i and node j, A i，j =0 means there is no edge between node i and node j.
5. The graph data preprocessing method according to claim 3, characterized in that: The GCN model is used to learn the subgraph embedding representation of node i, which is performed by the following formula: In the formula, k represents the kth node of the subgraph, K represents the number of subgraph nodes, and H k represents the representation of the k-th node.
6. The graph data preprocessing method according to claim 2, characterized in that: The embedded representation of the node i is obtained by the following formula: In the formula, h i (l-1) represents the representation of node i at layer l-1, W (l-1) represents the parameters of the l-1 layer, l represents the number of layers, is the ReLU activation function, h i Table 1 shows the overall embedding representation of the subgraph of node i.
7. The graph data preprocessing method according to claim 1, characterized in that: The graph contamination degree is: GPD(θ)=I(G,f θ (g))-I(G′,f θ (G′)), In the formula, G is the rough image, G' is the image to be predicted, I(G,f θ (G)) is G and f θ (G) The mutual information between θ (G) is the spatial representation of the rough graph, I(G', f θ (G')) is G' and f θ (G'), f θ (G') is the spatial representation of the image to be predicted, and GPD(θ) represents the difference between the mutual information of the coarse image and the image to be predicted.
8. A graph data preprocessing system, characterized in that: include: A rough graph acquisition module is used to remove multiple disturbance edges of the graph data to be processed to obtain a rough graph; The graph purification module randomly selects some edges from multiple perturbed edges and adds them to the rough graph to generate the graph to be predicted; Use the deep graph mutual information model DGI to learn the mutual information between the coarse graph and the spatial representation of the coarse graph, and to learn the mutual information between the graph to be predicted and the spatial representation of the graph to be predicted; The graph contamination degree is used to determine whether the difference between the two mutual information is less than a threshold. If not, multiple edges are selected from the remaining perturbation edges and added to the graph to be predicted to generate a new graph to be predicted, and it is determined again whether the difference between the mutual information of the new graph to be predicted and its spatial representation and the mutual information between the coarse graph and the spatial representation of the coarse graph is less than a set threshold. If so, a clean graph is obtained.
9. A computer device, characterized in that: It comprises a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the graph data preprocessing method according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the graph data preprocessing method according to any one of claims 1 to 7.